implement #107

Author: previ

.gitignore

@@ -38,6 +38,10 @@ photos/*.mp4
 photos/*.h264
 photos/*.json
 
+#cnn models
+cnn_models/*
+cnn_models/cache/*
+
 # Sounds recorded
 sounds/*.wav
 

camera.py

@@ -84,11 +84,11 @@ def __init__(self):
                         self._photos.append({'name': filename})
             self.save_photo_metadata()
 
-        #self._cnn_classifiers = {}
-        #cnn_model = config.Config.get().get("cnn_default_model", "")
-        #if cnn_model != "":
-        #    self._cnn_classifiers[cnn_model] = CNNManager.get_instance().load_model(cnn_model)
-        #    self._cnn_classifier_default = self._cnn_classifiers[cnn_model]
+        self._cnn_classifiers = {}
+        cnn_model = config.Config.get().get("cnn_default_model", "")
+        if cnn_model != "":
+            self._cnn_classifiers[cnn_model] = CNNManager.get_instance().load_model(cnn_model)
+            self._cnn_classifier_default = self._cnn_classifiers[cnn_model]
 
         self._camera.grab_start()
         self._image_cv = self.get_image()

cnn_classifier.py

@@ -27,6 +27,7 @@
 import time
 import logging
 
+import operator
 import numpy as np
 import tensorflow as tf
 
@@ -137,10 +138,7 @@ def classify_image(self,
         #logging.info( "time.cls: " + str(time.time() - s_t))
 
         results = np.squeeze(results)
-
-        pairs = {}
-        for i in results.argsort():
-            pairs[self._labels[i]] = results[i]
-
-        #logging.info(pairs)
+        result = results.argmax()
+        pairs = {self._labels[result]: results[result]}
+        logging.info(pairs)
         return pairs

cnn_manager.py

@@ -93,7 +93,7 @@ def train_new_model(self,
 
     def save_model_status(self, model_name, architecture, status):
         model_info = architecture.split("_")
-        self._models[model_name] = {"status": status, "image_height": model_info[2], "image_width": model_info[2]}
+        self._models[model_name] = {"status": status, "image_height": model_info[3], "image_width": model_info[3], "output_layer": "final_result"}
         self._save_model_meta()
 
     def wait_train_jobs(self):
@@ -105,6 +105,7 @@ def load_model(self, model_name):
         if model_info:
             return CNNClassifier(model_file = MODEL_PATH + "/" + model_name + ".pb",
                                  label_file = MODEL_PATH + "/" + model_name + ".txt",
+                                 output_layer=model_info["output_layer"],
                                  input_height = int(model_info["image_height"]),
                                  input_width = int(model_info["image_width"]))
 

cnn_models/models.json

@@ -1 +1 @@
-{}
+{"generic_fast_low":{"status":1.0, "image_height": "128", "image_width":"128", "output_layer": "MobilenetV2/Predictions/Reshape_1"}, "generic_slow_high":{"status":1.0, "image_height":"224", "image_width": "224", "output_layer": "MobilenetV2/Predictions/Reshape_1"}}

cnn_train.py

@@ -142,7 +142,7 @@ def __init__(self, manager, architecture):
         self.validation_batch_size = 100
         self.print_misclassified_test_images = False
         self.bottleneck_dir = "/tmp/bottleneck"
-        self.model_dir = "/tmp/imagenet"
+        self.model_dir = "./cnn_models/cache"
         self.final_tensor_name = "final_result"
         self.write_logs = False
 
@@ -160,7 +160,7 @@ def __init__(self, manager, architecture):
             raise Exception("Did not recognize architecture flag'")
 
         # Set up the pre-trained graph.
-        self.maybe_download_and_extract(self.model_info['data_url'])
+        self.maybe_download_and_extract(self.model_info['data_url'], self.model_info['model_dir_name'])
         self.graph, self.bottleneck_tensor, self.resized_image_tensor = (
           self.create_model_graph(self.model_info))
 
@@ -517,7 +517,7 @@ def run_bottleneck_on_image(self, sess, image_data, image_data_tensor,
         return bottleneck_values
 
 
-    def maybe_download_and_extract(self, data_url):
+    def maybe_download_and_extract(self, data_url, model_dir_name):
         """Download and extract model tar file.
 
         If the pretrained model we're using doesn't already exist, this function
@@ -526,7 +526,7 @@ def maybe_download_and_extract(self, data_url):
         Args:
           data_url: Web location of the tar file containing the pretrained model.
         """
-        dest_directory = self.model_dir
+        dest_directory = os.path.join(self.model_dir, model_dir_name)
         if not os.path.exists(dest_directory):
             os.makedirs(dest_directory)
         filename = data_url.split('/')[-1]
@@ -538,11 +538,10 @@ def _progress(count, block_size, total_size):
                                  (filename,
                                   float(count * block_size) / float(total_size) * 100.0))
                 sys.stdout.flush()
-
             filepath, _ = urllib.request.urlretrieve(data_url, filepath, _progress)
             print()
             statinfo = os.stat(filepath)
-            tf.logging.info('Successfully downloaded', filename, statinfo.st_size,
+            tf.logging.info('Successfully downloaded %s %d', filename, statinfo.st_size,
                             'bytes.')
         tarfile.open(filepath, 'r:gz').extractall(dest_directory)
 
@@ -1084,47 +1083,64 @@ def create_model_info(self, architecture):
                 tf.logging.error("Couldn't understand architecture name '%s'",
                                  architecture)
                 return None
-            version_string = parts[1]
+            v_string = parts[1]
+            version_string = parts[2]
             if (version_string != '1.0' and version_string != '0.75' and
-                version_string != '0.50' and version_string != '0.25'):
+                version_string != '0.50' and version_string != '0.5' and 
+                version_string != '0.35' and version_string != '0.25'):
                 tf.logging.error(
-                    """"The Mobilenet version should be '1.0', '0.75', '0.50', or '0.25',
+                    """"The Mobilenet version should be '1.0', '0.75', '0.50', '0.35', or '0.25',
             but found '%s' for architecture '%s'""",
                     version_string, architecture)
                 return None
-            size_string = parts[2]
+            size_string = parts[3]
             if (size_string != '224' and size_string != '192' and
                 size_string != '160' and size_string != '128'):
                 tf.logging.error(
                     """The Mobilenet input size should be '224', '192', '160', or '128',
            but found '%s' for architecture '%s'""",
                     size_string, architecture)
                 return None
-            if len(parts) == 3:
+            if len(parts) == 4:
                 is_quantized = False
             else:
-                if parts[3] != 'quantized':
+                if parts[4] != 'quantized':
                     tf.logging.error(
                         "Couldn't understand architecture suffix '%s' for '%s'", parts[3],
                         architecture)
                     return None
                 is_quantized = True
-            data_url = 'http://download.tensorflow.org/models/mobilenet_v1_'
-            data_url += version_string + '_' + size_string + '_frozen.tgz'
-            bottleneck_tensor_name = 'MobilenetV1/Predictions/Reshape:0'
+            data_url = 'http://'
+            model_file_name = None
+            bottleneck_tensor_name = None 
+            if architecture.startswith('mobilenet_v1'): 
+               data_url += 'download.tensorflow.org/models/mobilenet_v1_'
+               data_url += version_string + '_' + size_string + '_frozen.tgz'
+               bottleneck_tensor_name = 'MobilenetV1/Predictions/Reshape:0'
+               if is_quantized:
+                   model_base_name = 'quantized_graph.pb'
+               else:
+                   model_base_name = 'frozen_graph.pb'
+               model_dir_name = 'mobilenet_v1_'
+               model_dir_name += version_string + '_' + size_string
+               model_file_name = os.path.join(model_dir_name, model_base_name)
+               model_dir_name = '' 
+            else:
+               data_url += 'storage.googleapis.com/mobilenet_v2/checkpoints/mobilenet_v2_'
+               data_url += version_string + '_' + size_string + '.tgz'
+               bottleneck_tensor_name = 'MobilenetV2/Predictions/Reshape:0' 
+               model_dir_name = 'mobilenet_v2_'
+               model_dir_name += version_string + '_' + size_string
+               model_base_name = model_dir_name + '_frozen.pb'
+               model_file_name = os.path.join(model_dir_name, model_base_name)
             bottleneck_tensor_size = 1001
             input_width = int(size_string)
             input_height = int(size_string)
             input_depth = 3
             resized_input_tensor_name = 'input:0'
-            if is_quantized:
-                model_base_name = 'quantized_graph.pb'
-            else:
-                model_base_name = 'frozen_graph.pb'
-            model_dir_name = 'mobilenet_v1_' + version_string + '_' + size_string
-            model_file_name = os.path.join(model_dir_name, model_base_name)
             input_mean = 127.5
             input_std = 127.5
+            print(data_url)
         else:
             tf.logging.error("Couldn't understand architecture name '%s'", architecture)
             raise ValueError('Unknown architecture', architecture)
@@ -1138,6 +1154,7 @@ def create_model_info(self, architecture):
             'input_depth': input_depth,
             'resized_input_tensor_name': resized_input_tensor_name,
             'model_file_name': model_file_name,
+            'model_dir_name': model_dir_name,
             'input_mean': input_mean,
             'input_std': input_std,
         }
@@ -1170,11 +1187,3 @@ def add_jpeg_decoding(self, input_width, input_height, input_depth, input_mean,
         mul_image = tf.multiply(offset_image, 1.0 / input_std)
         return jpeg_data, mul_image
 
-
-if __name__ == '__main__':
-    cnn_trainer = CNNTrainer("mobilenet_0.50_128")
-    cnn_trainer.retrain(
-                image_dir="/home/pi/tensorflow/data/applekiwi",
-                output_graph="./cnn_models/applewiki_0_5_128.pb",
-                training_steps=10,
-                learning_rate=0.1)

main.py

@@ -34,7 +34,7 @@
 
 # Logging configuration
 logger = logging.getLogger()
-logger.setLevel(logging.WARNING)
+logger.setLevel(logging.DEBUG)
 sh = logging.StreamHandler()
 fh = logging.handlers.RotatingFileHandler('./logs/coderbot.log', maxBytes=1000000, backupCount=5)
 formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
@@ -96,7 +96,7 @@ def handle_home():
                            config=app.bot_config,
                            program_level=app.bot_config.get("prog_level", "std"),
                            cam=cam != None,
-                           cnn_model_names = json.dumps({}))
+                           cnn_model_names = json.dumps([[name] for name in cnn.get_models().keys()]))
 
 @babel.localeselector
 def get_locale():

photos/metadata.json

@@ -1 +1 @@
-[]
+[{"name": "DSC2.jpg"}]

scripts/interfaces_ap

@@ -0,0 +1,10 @@
+auto lo
+allow-hotplug wlan0
+iface lo inet loopback
+iface eth0 inet dhcp
+iface wlan0 inet static
+address 10.0.0.1
+netmask 255.255.255.0
+wireless-power off
+network 10.0.0.0
+broadcast 10.0.0.255

templates/config.html

@@ -251,14 +251,14 @@ <h3>{% trans %}Train new CNN Model{% endtrans %}</h3>
      <input type="text" id="i_cnn_image_tags" name="cnn_image_tags" value="">
      <label for="i_cnn_model_arch">{% trans %}Model architecture{% endtrans %}</label>
      <select name="cnn_model_arch" id="i_cnn_model_arch" value="">
-       <option value="mobilenet_1.0_128">MobileNet 1.0 128</option>
-       <option value="mobilenet_0.75_128">MobileNet 0.75 128</option>
-       <option value="mobilenet_0.50_128">MobileNet 0.50 128</option>
-       <option value="mobilenet_0.25_128">MobileNet 0.25 128</option>
-       <option value="mobilenet_1.0_224">MobileNet 1.0 224</option>
-       <option value="mobilenet_0.75_224">MobileNet 0.75 224</option>
-       <option value="mobilenet_0.50_224">MobileNet 0.50 224</option>
-       <option value="mobilenet_0.25_224">MobileNet 0.25 224</option>
+       <option value="mobilenet_v2_1.0_128">MobileNet v2 1.0 128</option>
+       <option value="mobilenet_v2_0.35_128">MobileNet v2 0.35 128</option>
+       <option value="mobilenet_v2_1.0_224">MobileNet v2 1.0 224</option>
+       <option value="mobilenet_v2_0.35_224">MobileNet v2 0.35 224</option>
+       <option value="mobilenet_1.0_128">MobileNet v1 1.0 128</option>
+       <option value="mobilenet_0.25_128">MobileNet v1 0.25 128</option>
+       <option value="mobilenet_1.0_224">MobileNet v1 1.0 224</option>
+       <option value="mobilenet_0.25_224">MobileNet v1 0.25 224</option>
      </select>
      <label for="i_cnn_model_train_steps">{% trans %}Training steps{% endtrans %}</label>
      <input type="range" id="i_cnn_train_steps" name="cnn_train_steps" min="10" max="100" step="10" value="50">

test/cnn_test.py

@@ -23,15 +23,15 @@ def test_train_set_1(self):
         cam = camera.Camera.get_instance()
         cnn = cnn_manager.CNNManager.get_instance()
         cnn.train_new_model(model_name=name,
-                            architecture="mobilenet_0.25_128",
+                            architecture="mobilenet_v1_0.50_128",
                             image_tags=["other","tomato","apple","kiwi"],
                             photos_meta=cam.get_photo_list(),
                             training_steps=20,
                             learning_rate=0.1)
         start = time.time()
         while True:
             model_status = cnn.get_models().get(name, {"status": 0})
-            if model_status.get("status") == 1 or time.time() - start > 600:
+            if model_status.get("status") == 1 or time.time() - start > 6000:
                 break
             print("status: " + str(model_status["status"]))
             time.sleep(1)
@@ -48,3 +48,32 @@ def test_train_set_1(self):
         cnn = cnn_manager.CNNManager.get_instance()
         cnn.delete_model(name)
         self.assertTrue(cnn.get_models().get(name) is None)
+
+    def test_train_set_2(self):
+        name="test_model_2"
+        cam = camera.Camera.get_instance()
+        cnn = cnn_manager.CNNManager.get_instance()
+        cnn.train_new_model(model_name=name,
+                            architecture="mobilenet_v2_0.5_128",
+                            image_tags=["other","tomato","apple","kiwi"],
+                            photos_meta=cam.get_photo_list(),
+                            training_steps=20,
+                            learning_rate=0.1)
+        start = time.time()
+        while True:
+            model_status = cnn.get_models().get(name, {"status": 0})
+            if model_status.get("status") == 1 or time.time() - start > 6000:
+                break
+            print("status: " + str(model_status["status"]))
+            time.sleep(1)
+        self.assertTrue(cnn.get_models().get(name).get("status") == 1.0)
+
+        cnn = cnn_manager.CNNManager.get_instance()
+        mod = cnn.load_model(name)
+        result = mod.classify_image("photos/DSC86.jpg")
+        print("result: " + str(result))
+        self.assertTrue(result["kiwi"] >= 0.9)
+
+        cnn = cnn_manager.CNNManager.get_instance()
+        cnn.delete_model(name)
+        self.assertTrue(cnn.get_models().get(name) is None)